GRAPH FEATURE ENGINEERING AND COORDINATE-BASED LEARNING FOR TRANSFERABLE AND ENERGY-EFFICIENT ARTIFICIAL INTELLIGENCE

Abstract

A comprehensive framework for efficient and scalable graph representation learning is presented, emphasizing coordinate-based and explicit structural methods. The research addresses the limitations of Graph Neural Networks (GNNs) in resource-constrained environments, including edge devices and large-scale deployments, by developing lightweight, non-neural alternatives. The first contribution is the Network Feature Embedding (NFE) pipeline, which integrates diffusion-based, positional, and structural descriptors into a unified representation for node classification. The second contribution is the Topology Coordinate-Driven Random Forests (TC-DRF) framework, which combines anchor-based topology coordinates with Random Forest classifiers for graph-level learning and cross-dataset transfer. Extensive evaluations of NFE and TC-DRF on vision, molecular, and social graph benchmarks demonstrate competitive predictive performance while substantially reducing computational overhead, memory footprint, and energy consumption. The proposed frameworks enable zero-shot cross-dataset transfer, maintain robustness under class imbalance, and support practical deployment in Green AI settings. Edge-device experiments, including deployment on Raspberry Pi hardware,confirm sub-millisecond inference latency and ultra-low energy usage. This research challenges the prevailing reliance on deep message-passing architectures for graph learning, demonstrating that explicit structural representations coupled with lightweight models provide viable, interpretable, and resource-efficient alternatives. The findings contribute to the advancement of scalable and sustainable graph learning methodologies and establish a foundation for future work in structural embeddings, dynamic graph analysis, and hybrid structural attributelearning models.